gh-xxxx: enable BOLT optimization of libpython

Before, we only supported running BOLT on the main `python` binary.
If a shared library was in play, it wouldn't be optimized. That was
leaving a ton of optimization opportunities on the floor.

This commit adds support for running BOLT on libpython.

Functionality is disabled by default because BOLT asserts on LLVM 15,
which is the latest LLVM. I've built LLVM tip and it is able to
process libpython just fine. So it is known to work.

Author: indygreg

Makefile.pre.in

@@ -720,7 +720,7 @@ profile-pgo-apply-stamp: profile-pgo-analyze-stamp
 # passes.
 
 # List of binaries that BOLT runs on.
-BOLT_BINARIES = $(BUILDPYTHON)
+BOLT_BINARIES = @BOLT_BINARIES@
 
 BOLT_INSTRUMENT_FLAGS ?= @BOLT_INSTRUMENT_FLAGS@
 BOLT_APPLY_FLAGS ?= @BOLT_APPLY_FLAGS@
@@ -732,8 +732,6 @@ clean-bolt:
 	find . -name '*.bolt_inst' -exec rm -f {} ';'
 	# The data files they produce.
 	find . -name '*.fdata' -exec rm -f {} ';'
-	# Copied of binaries before BOLT application.
-	find . -name '*.prebolt' -exec rm -f {} ';'
 
 # BOLTs dependencies are a bit wonky.
 #
@@ -751,28 +749,36 @@ profile-bolt-prebuild-stamp: @MAKE_BOLT_NATIVE_DEPENDENCY@
 
 profile-bolt-instrument-stamp: profile-bolt-prebuild-stamp
 	for bin in $(BOLT_BINARIES); do \
-	    if [ -e "$${bin}.prebolt" ]; then \
-	        echo "Restoring pre-BOLT binary $${bin}.prebolt"; \
+	    prebolt="$${bin}.prebolt"; \
+	    if [ -e "$${prebolt}" ]; then \
+	        echo "Restoring pre-BOLT binary $${prebolt}"; \
 	        mv "$${bin}.prebolt" "$${bin}"; \
-	    fi \
+	    fi; \
+	    cp "$${bin}" "$${prebolt}"; \
 	done
 	# Ensure prior BOLT state is purged.
 	$(MAKE) clean-bolt
-	@LLVM_BOLT@ ./$(BUILDPYTHON) -instrument -instrumentation-file-append-pid -instrumentation-file=$(abspath $(BUILDPYTHON).bolt) -o $(BUILDPYTHON).bolt_inst $(BOLT_INSTRUMENT_FLAGS)
+	for bin in $(BOLT_BINARIES); do \
+	    @LLVM_BOLT@ $${bin} -instrument -instrumentation-file-append-pid -instrumentation-file=$(abspath $${bin}.bolt) -o $${bin}.bolt_inst $(BOLT_INSTRUMENT_FLAGS); \
+	    mv "$${bin}.bolt_inst" "$${bin}"; \
+    done
 	touch $@
 
 profile-bolt-run-stamp: profile-bolt-instrument-stamp
-	$(RUNSHARED) ./$(BUILDPYTHON).bolt_inst $(PROFILE_TASK) || true
+	$(RUNSHARED) ./$(BUILDPYTHON) $(PROFILE_TASK) || true
 	touch $@
 
 profile-bolt-analyze-stamp: profile-bolt-run-stamp
-	@MERGE_FDATA@ $(BUILDPYTHON).*.fdata > $(BUILDPYTHON).fdata
+	for bin in $(BOLT_BINARIES); do \
+	    @MERGE_FDATA@ $${bin}.*.fdata > $${bin}.fdata; \
+	done
 	touch $@
 
 profile-bolt-apply-stamp: profile-bolt-analyze-stamp
-	@LLVM_BOLT@ ./$(BUILDPYTHON) -o $(BUILDPYTHON).bolt -data=$(BUILDPYTHON).fdata $(BOLT_APPLY_FLAGS)
-	mv $(BUILDPYTHON) $(BUILDPYTHON).prebolt
-	mv $(BUILDPYTHON).bolt $(BUILDPYTHON)
+	for bin in $(BOLT_BINARIES); do \
+	    @LLVM_BOLT@ "$${bin}.prebolt" -o "$${bin}.bolt" -data="$${bin}.fdata" $(BOLT_APPLY_FLAGS); \
+	    mv "$${bin}.bolt" "$${bin}"; \
+	done
 	touch $@
 
 # End of profile-based optimization rules.

configure

@@ -2083,6 +2083,23 @@ if test "$Py_BOLT" = 'true' ; then
   fi
 fi
 
+# Enable BOLT optimizations of libpython. Optional for now due to known
+# crashes on LLVM 15. Seems to be fixed in LLVM 16.
+AC_SUBST(BOLT_BINARIES)
+BOLT_BINARIES='$(BUILDPYTHON)'
+
+AC_MSG_CHECKING(for --with-bolt-libpython)
+AC_ARG_WITH(bolt_libpython,
+            AS_HELP_STRING([--with-bolt-libpython], [enable BOLT optimization of libpython (WARNING: known to crash BOLT)]),
+            [with_bolt_libpython="yes"],
+            [with_bolt_libpython="no"])
+AC_MSG_RESULT($with_bolt_libpython)
+
+if test "${enable_shared}" = "yes" -a "${with_bolt_libpython}" = "yes"
+then
+  BOLT_BINARIES="${BOLT_BINARIES} \$(INSTSONAME)"
+fi
+
 AC_ARG_VAR(BOLT_INSTRUMENT_FLAGS, Arguments to llvm-bolt when instrumenting binaries)
 AC_MSG_CHECKING(BOLT_INSTRUMENT_FLAGS)
 if test -z "${BOLT_INSTRUMENT_FLAGS}"